Title: Observational Average Treatment Effect Estimation with Images as Confounders

Abstract: This talk lays out a straightforward framework for treatment effect inference with image-valuate control covariates, and numeric treatments and outcomes, a setting that is increasingly more common as image data becomes more available. The proposed method combines predictions from deep convolutional neural networks with semiparametric de-biasing strategies to maintain good statistical behavior of the resulting estimates. The proposed network architecture and training scheme improves upon several existing alternatives by combining standard regularization methods for image data with de-biasing regularization at training time. Simulation studies show that this approach enables the resulting model to achieve optimal performance both in finite samples and asymptotically. The proposed method is applied to the problem of estimating whether social media users would react differently to the same image being shared by a democratic or republican account, and find that images shared by republican accounts tend to, on average, receive more user interaction on social media.